Dissolved oxygen (DO) is an important parameter for water quality. Appropriate DO concentration range is helpful to the growth of aquatic products. Prediction the concentration change of DO is important for environment early warning. In order to predict DO concentration change accurately and quickly, a prediction model for DO concentration in aquaculture is proposed based on optimized echo state networks (ESN) in this work. The method of bidirectional construction of training samples is integrated with the ESN model to build a DO prediction model, which solves the problem of network free parameter determination in traditional ESN models and performance deterioration when the reserve pool size is large. It also solves the problem that DO cannot be predicted quickly and accurately in aquaculture. The test results show that the improved ESN has good robustness. At the same time, when the reserve pool is relatively large, the overfitting phenomenon of traditional ESN can be effectively reduced, and the generalization performance of the model is improved.